Müller et al. improved deep learning (DL) generation of retinal fundus images by overriding shortcut learning that typically occurs due to the presence of spurious correlations. DL models were developed to disentangle the 3 main factors that distinguish retinal fundus images from each other: style, attribute, and camera type. All models were trained with anonymized fundus images that had no disease and were cropped to the same dimensions. First, an encoder model was developed to encode features of the images and to map those attributes to their own subspaces. Disentanglement was reduced by penalizing shared information between subspaces. Additionally, a generative model was developed to generate realistic fundus images from disentangled data, using StyleGAN2 as its framework. Some limitations in the generative model included a difficulty in capturing fine details and disentangling certain variables that were likely due to inherent correlations present in the datasets used to train the model. However, the model can be feasibly trained with larger datasets to improve its disentangling ability. Tests were performed with both the encoder and the generative models to assess sensitivity and quality of generated images. Quantitative and qualitative analysis based on these tests demonstrated that images produced by the generative model were very similar to the quality of real fundus images.